Árvore de Regressão aplicada à identificação e priorização de perdas não técnicas

Abstract

This dissertation presents a target prioritization methodology for the inspection of non-technical losses (commercial losses) in electricity distribution systems, based on predictive modeling with regression trees. The main goal is to identify consumer units with higher propensity to fraud and consumption irregularities, in order to support normalization actions and mitigate the technical, economic, and social impacts of such losses. A bibliographic and bibliometric review of machine learning methods applied to non-technical loss detection was conducted, along with an indepth study of regression tree techniques and their applicability in this domain. The proposed model was developed using real data from a distribution company, including consumption history, consumer attributes, and operational records, and comprised database preparation, selection of predictive variables, construction and cost-complexity pruning of regression trees, and predictive performance evaluation. Results highlight the effectiveness of the model in prioritizing consumer units with the greatest impact on non-technical losses, reinforcing its potential as a decisionsupport tool in the electricity sector.